RD-VIO: Robust Visual-Inertial Odometry for Mobile Augmented Reality in Dynamic Environments.

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Visual-inertial odometry (VIO) systems struggle with dynamic environments and pure rotational motion.
• Robust keypoint detection and matching are crucial for accurate VIO.

Purpose of the Study:

• To develop a novel VIO system, RD-VIO, capable of handling both dynamic scenes and pure rotational motion.
• To improve the robustness and accuracy of VIO in challenging real-world scenarios.

Main Methods:

• Proposed an IMU-PARSAC algorithm for robust two-stage keypoint detection and matching using visual and IMU data.
• Developed a motion type detection and deferred-triangulation technique to manage pure rotational motion.
• Integrated pure-rotational frames as special subframes providing additional constraints during bundle adjustment.

Main Results:

• The IMU-PARSAC algorithm demonstrated robust landmark matching.
• The deferred-triangulation technique effectively handled pure rotational motion.
• RD-VIO showed significant advantages over other VIO methods in dynamic environments during evaluations on public datasets.

Conclusions:

• RD-VIO offers a robust solution for visual-inertial odometry in dynamic scenes and during pure rotations.
• The proposed methods enhance the reliability and accuracy of VIO systems.
• RD-VIO presents a promising advancement for applications requiring precise motion tracking.